01 / AI ENRICHMENTPreserve the source beside every enrichment
Keep original content and generated interpretation in different fields. A label or summary should carry the source item and revision it depends on. When a reviewer disagrees, the underlying evidence should still be available. This separation also lets you improve a task without pretending that the publisher revised its own material.
02 / AI ENRICHMENTValidate generated output before publishing
Define a small, explicit output contract before asking for a result. A categorizer can use an approved vocabulary and a review state; a summarizer can preserve uncertainty and identify missing evidence. Validate the shape, then evaluate whether the result is actually supported. These checks answer different questions and should not collapse into one success flag.
03 / AI ENRICHMENTTreat retrieved content as untrusted data
External feed text should not control your application’s permissions. Keep task instructions in trusted configuration and limit the enrichment worker to the operations it needs. A summary task can produce a draft without access to a publishing tool. Add capabilities only after the workflow has a clear reason and a review path.
Suggest one label, then evaluate it
Choose a narrow categorization task and a small set of representative items. Keep the source content visible beside the proposed label, including a route to leave the item unassigned. Compare suggestions with an agreed review standard and record the error types that matter. Only expand into summaries or more autonomous behavior after the initial task provides a measurable benefit to the people using the queue.